Confessions can keep language models honest

4 points| xavierlint | 2 months ago |openai.com